Velde: constructing cell potential landscapes by RNA velocity vector field decomposition

Junbo Jia, Luonan ChenarXiv 2023 · DOI · arXiv

TL;DR

Velde is a new computational framework that uses vector field decomposition to untangle two fundamental biological processes: cell differentiation and the cell cycle. By applying the Helmholtz-Hodge decomposition to RNA velocity data, it simultaneously constructs two distinct "potential landscapes," allowing researchers to visualize how cells both "roll down" developmental hills and "spin" through their reproductive cycles.

What Problem Does This Paper Address?

Since 1957, biologists have used the "Waddington landscape" as a metaphor for cell development: imagine a ball (a cell) rolling down a complex terrain of hills and valleys. The valleys represent stable cell types, and the paths represent differentiation.

While powerful, this metaphor is incomplete. In reality, cells aren't just moving toward a final state; they are often simultaneously trapped in a "cell cycle"—a repeating loop of growth and division. Existing computational methods for building these landscapes usually focus solely on the "downhill" differentiation part, effectively ignoring the "spinning" motion of the cell cycle. This creates a flattened, potentially misleading view of cellular dynamics. The authors of this paper sought a way to mathematically separate these two concurrent processes to provide a high-fidelity map of cell behavior.

Methods

The authors introduce Velde (RNA Velocity vector field decomposition). The core engine of Velde is the natural Helmholtz-Hodge decomposition (nHHD), a mathematical technique traditionally used in fluid dynamics to break a complex flow into simpler parts.

The workflow follows three main steps:

  1. Vector Field Construction: It takes RNA velocity data (which predicts the future state of individual cells) and projects it into a low-dimensional space like UMAP or t-SNE to create a continuous vector field.
  2. Decomposition: Using nHHD, Velde splits this field into three components: a gradient field (flow from source to sink), a rotation field (local spinning), and a harmonic field (flow influenced by external boundaries).
  3. Landscape Generation:
    • The gradient and harmonic fields are merged to form a curl-free component, which represents the "downhill" trek of differentiation.
  4. The rotation field becomes the curl component, representing the cell cycle.
  5. Finally, the tool calculates potential functions for both, resulting in two 3D landscapes.

Datasets

The authors validated Velde using a mix of controlled and biological data:

  • Synthetic Vector Fields: A "ground truth" test where a known gradient field and rotation field were combined to see if Velde could perfectly separate them (it did).
  • Synthetic scRNA-seq Data: Simulated single-cell data to test the pipeline's robustness.
  • Real scRNA-seq Datasets:
    • Dentategyrus: Mouse brain development data.
  • Human Hematopoiesis: The formation of blood cells.
  • Pancreatic Development: A complex dataset known to contain both strong differentiation signals and active cell cycling.

Key Insights and Ideas

  • The "Spinning Hill" Problem: The paper posits that cellular heterogeneity arises from two concurrent processes: transition (changing state) and cycle (rotating around a state). By treating these as mathematically distinct components of a single vector field, we gain a much clearer biological picture.
  • Potential as Energy: In the differentiation landscape, "high potential" areas represent progenitor cells with high plasticity, while "low potential" valleys represent mature, specialized cells.
  • The Conical Peak: A key "aha" moment occurs in the cell cycle landscape. While differentiation looks like a valley, a strong cell cycle appears as a steep, cone-shaped peak. The steeper the cone, the more "forceful" or active the cell cycle is in that specific population.

Key Results

  • Successful Separation: In the synthetic tests, Velde's reconstructed curl-free field was nearly identical to the hidden ground-truth gradient field, proving the math holds up.
  • Pancreas Discovery: In the pancreatic development dataset, Velde successfully identified a strong cell cycle at a specific location in the cell population. This was visualized as a distinct conical peak in the cycle landscape, whereas the brain and blood datasets showed almost no obvious cycle potential, matching biological expectations for those specific samples.
  • Consistency: The differentiation landscapes produced by Velde remained consistent with known developmental trajectories (e.g., progenitors at the top, mature cells at the bottom), proving that adding the cycle analysis doesn't "break" the traditional differentiation map.

Limitations and Future Work

The authors acknowledge two primary hurdles:

  1. RNA Velocity Sensitivity: Velde depends on the quality of the input RNA velocity. If the underlying velocity estimation is noisy or biased due to sequencing limitations, the resulting landscapes will be inaccurate.
  2. Embedding Topology: The choice of low-dimensional embedding (like UMAP) significantly affects the "aesthetics" and accuracy of the landscape. If the embedding doesn't capture the true high-dimensional topology of the cells, the decomposition might miss subtle biological signals.

Future Work: The team aims to move beyond visualization to extract specific gene regulatory mechanisms that drive the "slopes" and "spins" identified in their landscapes.

Summary updated

Abstract

The Waddington landscape serves as a metaphor illustrating the developmental process of cells, likening it to a small ball rolling down various trajectories into valleys. Constructing an epigenetic landscape of this nature aids in visualizing and gaining insights into cell differentiation. Development encompasses intricate processes involving both cell differentiation and cell cycles. However, current landscape methods solely focus on constructing a potential landscape for cell differentiation, neglecting the accompanying cell cycle. This paper introduces a novel method that simultaneously constructs two types of potential landscapes using single-cell RNA sequencing data. Specifically, it presents the natural Helmholtz-Hodge decomposition (nHHD) of a continuous vector field within a bounded domain in n-dimensional Euclidean space. This decomposition uniquely breaks down the vector field into a gradient field, a rotation field, and a harmonic field. Utilizing this approach, the RNA velocity vector field is separated into a curl-free component representing cell differentiation and a curl component representing the cell cycle. By calculating the corresponding potential functions, potential landscapes for both cell differentiation and the cell cycle are obtained. Finally, the efficacy of this method is demonstrated through its application to synthetic and real datasets.
BibTeX
@misc{jia2023velde,
  author = {Junbo Jia and Luonan Chen},
  title = {{Velde: constructing cell potential landscapes by RNA velocity vector field decomposition}},
  year = {2023},
  eprint = {2311.10403},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.10403},
}
Figure